Unleashing the potential of digital twins: a new era with aeronautics 4.0   
  

2.1 Aeronautics value chain analysis

The aerospace value chain consists of a series of processes, and digitalizing it begins with an assessment of these processes and their potential for integration into the fully integrated, three-dimensional Industry 4.0 perspective of the smart factory. Through Figure 1. we attempt to frame all the aspects of the value chain.

Figure 1. Aeronautics value chain break down and ecosystems overview.

The value chain is divided into three main parts:

Four main ecosystems are defined to frame the value chain of the aeronautics industry: development of engineering activities (study, design, research, and development), electrics and wiring (development of electrical systems and harnesses ….), assembly (mechanical equipment, electronic modules, wires and harnesses; structural parts …), and maintenance and repair operations (painting, component, and engine repair…). The mainstreaming of these improvements is achieved by leveraging Industry 4.0, technologies to support the realization, updating, and maturity enhancement of existing processes.

2.2 PESTEL analysis for Aero 4.0 integration

Tables 1, 2, 3, and 4 provide an overview of these tools and of the breaking points that we have defined as the value-added workstreams involved in this digital transition. The specification of these workstreams provides a clearer picture of the impact of digitalization on the value chain in the short and long terms. The representation of the current value chain in addition to the PESTEL analysis presented in Table 1, 2, 3, and 4 are used to define the targeted areas of improvement for each ecosystem in the selected business segment and find areas of added value for digital threads and digital twins.

2.3 D2TSAero – Digital Thread Twin for Smart Aeronautics

The proposed solution D2TSAero, represented in Figure 2, consists of three key network elements: suppliers and influencers, digital twins and digital threads, customers, and their customers.

Figure 2. D2TSAero – Digital Thread Twin for Smart Aeronautics.

Network of Suppliers and Influencers: The first network, namely that of suppliers and influencers, includes all parties with an influence on the operation of the main ecosystems, such as standardization or auditing organizations, which establish domain management procedures and define best practices; government; and its constituents, which manage the domain’s strategic development model and influence the local, national, and international organizational development of ecosystems.

Second, we identify all stakeholders involved in the production of the finished product and the supply of the main materials, elementary means for the development of the environment in which the finished product is produced, and the labor that completes this operation.

The field is classified as labor-intensive, and technological advances are reducing the number of complex or repetitive operations and helping the workforce improve its skills. The addition of universities and research organizations to this layer is intended to reinforce this aspect and thus continually update the sector’s various missing links for better integration. Each party in the network is represented by its AAS.

The AAS concept was introduced by the European Platform 4.0 coalition to ensure the deployment of Industry 4.0, within factories, and the implementation of the digital twin concept. It is the digital image of the entities within the virtual world, represented by the primary element encompassing the identification and domain information models of the entities, and the secondary elements constituted by the communication interfaces and management component enabling the various models and subsystems of the AAS to be managed.

Network of Digital twins and Digital Thread: The second network is that of the digital twins and digital value chain and covers the new development model for the various production environments and ecosystems involved in the complete manufacturing of the finished product.

Each ecosystem is represented by a global digital twin. Digital twin ecosystems communicate with each other and with the other two networks via dedicated application programming interfaces. The basic architecture of the digital twin ecosystems was developed according to the architecture shown in the figure and in compliance with the two standards ISO/IEC 30141:2018 for an Industrial Internet of Things reference architecture and ISO for the development of digital twins within manufacturing industries. The choice between these two standards is driven by our motivation to propose a generic and scalable architecture.

We have added a security layer to the base layer, made up of two actors: the first is that of intelligent contracts, developed on the basis of the description of network agents and their interactions, and monitored by intelligent encryption and artificial immunity algorithms; the second actor is based on Distributed Ledgers technology for managing exchanges and transactions between entities in the digital world. The interaction between these two elements makes it possible to build an immune interface between the different agents of the architecture given the criticality of the domain, despite the need to ensure horizontal and vertical integration between its players.

The basic twin architecture is composed of two layers: the digital twin layer and the service layer, which comprises three services: collaborative maintenance management, exploration of ecosystem entities, and quality management and continuous improvement.

The digital twin layer consists of twins and industrial entities as well as a workforce contributing to the creation of added value, its information perception and contextualization layer, and the intrinsic twin layer encompassing information models, physical entity simulation models, and twin functionality.

Cross domain functions: We also defined cross-domain functions for all parts of the architecture, such as security management, license management, and digital authorities. Communication between these different layers is ensured by communication protocols and mediating communication layers such as OPCUA, MT Connect, Automation ML, and OPCUA. The choice of protocol or integration depends on several criteria, including technical interoperability, cyber security, and deployment costs. The integration of standard data formats and information models within the architecture for semantic and syntactic interoperability is ensured by AAS directories developed based on an agile and collaborative engineering process provided by twin-type simulation, cloud platforms, and virtual prototyping.

Within the digital value chain model, logistics and reverse engineering are integral to the ecosystem development process. The digital twin’s reactivity to waste through optimization functionality enables the management of the non-value-added produced by each process. This non-value-added is communicated to the collaborative maintenance and continuous improvement departments for in-depth analysis, supported by the network of decision-makers. All modifications and improvements produced by the synergy of artificial and human intelligence were communicated to Ecosystem 1 and integrated into the new instance model of the twin types, thus facilitating reverse engineering and digitizing the process of testing new structures for improved performance. Meanwhile, waste management is handled by process twins, which integrate the reverse value chain into the overall ecosystem management.

Network of Customers and their customers: The last layer of the architecture is that of customers and customers, represented by airlines, international manufacturers, and end-users who are customers of airlines. From a Lean management perspective, the aim is to create an interactive, dynamic interface with this network via the services in order to manage the entire development and engineering process in an optimized, end-to-end way, but also to offer this network the possibility of benefiting from the integration of digitalization tools within the value chain for triple synchronization of costs, deadlines, and quality.

3.1 Decision Making Problem context description

The simulation is based on the model of an aircraft distribution system proposed by Ref. 28 for the monitoring of safety-related condition indicators, as shown in Figure 3(a). The choice of this particular use case was motivated by the first area of concern discussed previously, which extended the proposed framework discussed by the different developed DT instances and types. The remainder of the use case is based on the analysis of the dataset proposed for this system, which is considered to be produced data from the real system.

Figure 3. Real (a) and Virtual twin for Fuel distribution system (b).

The proposed system is inspired by the system described in previous sections and is composed of three main components: tanks, five main pumps and valves, and a refuelling point. The central tank is connected to two valves that control the amount of fuel that flows to each of the two engine lines. The lines are connected to the front and rear engines by a main pump and a valve that controls, in addition to the pump, the flow of fuel through the engine. Each line is connected to a main tank provided with fuel from the central tank by means of a main pump and a valve for flow control. Different flow, pressure, temperature, and level sensors are installed to reinforce system safety and report measurements for flight control and for further analysis. A refuelling point is activated when the level of fuel crosses a minimum within the central tank.

DT instances have the main mission of adapting system behavior to environmental changes by optimizing individual assets and collective performances. The proposed system consists of a set of complex systems that exhibit particular properties and models that relate to the areas of interest for decision makers and users’ operational and business teams. The two decision-making problems interact to achieve the two main defined sub-goals for DT agents: smart maintenance scheduling and control optimization. Decision-making problems are defined through States, Actions, Rewards and State-Action Transition matrices. In the context of this paper, we focus on defining smart maintenance scheduled through the proposed solution. Operation sequence control optimization will be included through the rewarding system and detailed in further work.

DT Types work in parallel to help the instances perform best. DT Types continuously learn to reinforce their instances’ decision-making capabilities. For this purpose, we developed a simulated environment that helps simulate faulty and healthy scenarios from previous experience, but also results from random noise and conflicting situations. Real scenarios exploited from the proposed dataset for the use case were augmented with a newly developed dataset that aimed to integrate security threats affecting the three physical, cyber, and networking parts of each asset, particularly for the measurements and control system. As highlighted previously, we want to explore the situation of coordinated learning to resolve conflicting goals by integrating both security and safety concerns. Coordination completes the adaptive and smart monitoring loop of the system by evaluating the correlations between safety and security risks for implemented safety equipment and measurement systems, such as sensors and valves.

3.2 DT Types models

The model was developed on the basis of the Pump Head and torque characteristic curve data interpolation according to the empirical model of the head and torque and pump Affinity Laws based on Eq. (1) and Eq. (2), respectively.

$\mathit{Hp},Q,w$: head, Flow rate and speed for pump

$Q_0,P_0,H_0$: nominal head, Flow rate and speed for pump

Coefficient approximation was performed on MATLAB Curve Fitting Tool according to selected operational points from the chose centrifugal pump for nominal speed:

The dynamics of the electrical driver are modelled by speed variation across the shaft and pump torque evolution as a load for the motor represented by Eq. (3) and Eq. (4), respectively, and loss expressions are considered to reproduce real motor operation according to Eq. (5). The regulation pump speed through the diver was introduced by a PI regulator on the motor torque represented by Eq. (6).

$T_e$: Pump drive torque loss

$J_{∆}$: Intertia for pump drive

$D_{\omega }$: Damping factor

$T_p$: Pump torque

$Y\left(s\right)$: Process variable

$u\left(s\right)$: Control output

$W\left(s\right)$: Set point for control

To adapt the system to different operating scenarios, PID parameters were tuned according to a PSO Loop with system models of the plant. At each new goal plan for the agent, the PID function is called to provide agents with new optimized input variables. The activation of the Loop is Controller with a goal flowchart.

The dynamics of the tank are modeled by the tank volume, flow, and level evolution during the simulation according to Eq. (7), Eq. (8), Eq. (9) and Eq. (10). for level, fluid pressure, and flow, respectively. The flow at the outlet of the tank depends on the valve characteristics and the control signal to the valve-opening position. Eq. (11) describes the model tank level as an output variable for a control loop, which is discussed further for system fault modeling.

3.3 Field Interrogation, Optimization, Prediction and Learning

As Described earlier, the two DT networks of type and instance collaborate to achieve defined smart maintenance planning.

To achieve this goal, DT agents exploit their acquired capabilities and assigned roles by default. One of the main roles deployed by instance and type is Online and Offline learning.

The learning model for the two networks is defined by Reinforcement Learning RL quadruple $⟨S_t,A_t,T_s,R_t⟩$ agent states within the environment, Actions, Transition probabilities, and rewards.

Agent States within the environment are defined by three dimensions that describe the main factors that orient agents’ decisions and actions.

The first dimension concerns Highly Critical Assets HCA states that can be considered as the main influencers of the overall system performance.

The second dimension is defined through the Asset Health Index, which portrays asset status by the combination of main asset deterioration and health evolution models and the impact of modifiers on this status. Three modifiers were considered: health, reliability, and operational modifiers.

The third dimension concerns failure evolution, which we define as the failure rate that can be linked to both deterioration modes and health index evolution to reduce the state-space dimension.

Inspired by Ref. 29, we define critical asset states as New, Deteriorating, Deteriorating critical, under PM, available to adjust, and failed. Asset’s criticality code that will be included in the Functional Safety Sub model will be assigned to assets based on a multicriteria analysis of asset health deterioration impact degrees, and this will be detailed throughout our further works. Several health index evaluation models have been proposed in the literature; for the purpose of this paper, and in order to reduce problem definition complexity referring to asset performance and deterioration modes, we define the Asset Health Index (AHI) $H_i^t$ through flow and efficiencies, as proposed in the framework developed by.³⁰ The choice of this framework is motivated by our interest in integrating different modifier factor impacts on asset performance and index-wide use in the aviation domain. The explicit representation of the states according to the health index and failure rates is given by Eq. (12) , Eq. (13) and Eq. (14). The failure rate in the context of this study was assumed to follow a probabilistic Weibull Distribution.

$S_i$: Asset state i at instant t.

$F_R^t$, $F_E^t$,, $F_H^t$, $F_L^t$: Health, load, reliability, and environmental factors

${\lambda }_j\left(t\right)$: Failure rate for asset i at time t.

${\beta }_{i,1}^d$: Weibull distribution slope for Asset I.

${\eta }_{i,2}^d$: Weibull distribution scale for Asset I.

To evaluate their performance in addition to asset state transitions, DT networks receive a reward at each decision epoch that helps them progress with their action plans through their operation.

Agent rewards are defined by the maintenance policy value trajectories. The agents’ decision epochs are defined by several decision episodes at each decision episode agent across various states that are characterized by their values and impacts on maintenance value trajectory optimization. The usage costs of an asset are defined by Eq. (15). $C_O,C_{energy},C_{m }\text{and} C_S$ represent the operation, energy, maintenance, and out-of-service costs, respectively, in Eq. (16), Eq. (17), Eq. (18) and Eq. (19). This cost was subtracted from the asset profit value to constitute the reward function represented in Eq. (20).

These costs are defined as the cumulative usage phase costs and reward for each agent within the network for each decision epoch. All usage-phase costs are considered in this case study to apprehend the benefits of the proposed architecture. The DT agent reward is defined by Eq. (21). Safety and satisfaction functions are represented by Eq. (22) and Eq. (23).

Within the cost function, we define three coefficients for preventive, corrective, and repair, and test costs ${\mathit{a}}_{\mathit{k}}$, ${\mathit{b}}_{\mathit{k}}$, ${\mathit{c}}_{\mathit{k}}$ respectively.

Action space consists of five main actions and their derivatives, which are defined through ISO 14624 definitions of major maintenance activities. Two categories are defined: the first includes one action, which is No Maintenance performed, and the second category, which focuses on different maintenance actions, includes four maintenance activities.

Check and adjust or Minor Maintenance, which consists of the execution of different inspection operations while maintaining asset operation; the second maintenance action within this category is Preventive Maintenance, which includes the set of planned checks for preventing asset operation and the adjustment of working parameters, the third action includes Corrective Maintenance actions that result from asset failure during its operation as a critical result for agent control. CM actions are defined according to the resulting failure modes, its causes, and effects, and their definition by architecture agents’ networks will be detailed further. The last action within this category is replacement, which transits assets to a new state and is considered the worst case for DT Types and instances network performance evaluation.

Algorithm 1 represents the agent’s capability for anomaly prediction and field interrogation based on the proposed model for Health Index estimation. The proposed algorithm processes time-series data related to asset conditions, extracts features, and uses a classifier to identify anomalies or normal behavior for each asset. It updates the failure probability and behavior label accordingly, and computes certain metrics based on the extracted data.

Finally, it adds the computed rule to a set of behavioral rules and archives the scenario for further analysis.

Classification is performed based on a model pretrained by DT Types using a case base of healthy and faulty simulation results.

Algorithm 2 DT type capability for smart maintenance planning and action optimization based on RL logic.

Algorithm 3 shows the DT-Instances Learning algorithm for Proactive Decision Making with the integration of collaborative learning between agents through the exploration of similar cases based on the agent’s own experiences and peers’ experiences.

In the context of the proposed use case, five runs with five operating points for each run were simulated to represent healthy conditions of the system. Faulty conditions were simulated using MATLAB by injecting faults at each epoch and their combination through the introduction of a Failure_Scenarios_Simulation () function that can be manipulated by system users to test, simulate, and generate ensemble fault data stores. Table 5 presents the experimental details for the heathy scenarios with different loads. We assume the initial condition that assets have been operating for 100h with 20 start-ups of the Front and Rear Pumps. The estimated flight time is considered to estimate the operational time of assets for the modifier estimation.

Three main critical failure modes for assets, as well as one incipient and intermittent failure mode, were simulated to test different scenarios. Faults are modelled within MATLAB/SIMULINK using different fault models defined by StartTime, Duration, Fault_Value_Set, and Fault_Sevirity_Category. This section introduces Failure_Scenarios_Simulation () in MATLAB for the discussed use case. The fault mode sets are represented by different gains from Eq. (24)–(30), respectively.

Table 6 presents the scenarios for the faulty behavior of the pump. Security faults arising from security threats are also considered.

The Weibull parameters estimated from a bibliographic review of aircraft fuel distribution reliability data enable the estimation of the probability parameters for different failure modes of the driven pumps within the system and their subunits. According to the Table above, ${\alpha }_1$ and ${\beta }_1$, ${\alpha }_2$ and ${\beta }_2$ are represented for the early and wear-out life phases of the pump unit Eq. (31) to Eq. (34).

The proposed reward function of the system was calculated based on the estimated maintenance costs of each subunit and its failure modes. The PM, CM, and RM costs for the subunit discussed in the context of this paper are presented in Table 8, and the failure mode costs and their respective necessary PM, CM, and RM duration are presented in Tables 7 and 9.

An example snapshot of the standard fault definition document for the fault submodel of digital twins in JSON format according to the standard definition of fault properties is presented in Table 10. The proposed type of fault model aims to fulfil the concept from digital thread to have a standard and sharable definition of maintenance main concepts that can be used by the DT network but also its stakeholders.